Deep Learning Architectures and Their Efficacy in Surveillance Face Recognition: A Study with the ChokePoint Dataset

This work investigates the efficiency of various deep learning architectures in the field of surveillance face recognition using the ChokePoint Dataset. The focus is on person identification through facial images, leveraging state-of-the-art deep learning techniques. Our methodology centered on evaluating deep learning architectures and tailoring experiments around criteria such as image resolutions and training epochs. For the experiments, two image sizes were considered: 75 × 75 pixels and 224 × 224 pixels. This work used the ChokePoint Dataset, chosen for its real-world surveillance attributes. All models were trained across a consistent epoch number of 300. Investigating various architectures revealed that simpler models like MobileNet thrive at lower resolutions, whereas complex models favor larger images. Beyond mere accuracy, factors like computational efficiency and real-time processing are vital. MobileNet achieved a remarkable 93.45% test accuracy. Its architectural advantages, such as depth-wise separable convolutions, allow for reduced computational demand while preserving essential feature extraction capabilities. This blend of high accuracy, efficiency, and adaptability to smaller images makes MobileNet an optimal choice for person identification from facial images. In conclusion, our experiments highlight the impressive balance between performance and efficiency achieved by MobileNet.